In promoting the associative learning account of mirror neurons (MNs), Cook et al. have chosen to adopt an extreme version. Not only have they argued that the matching properties of MNs (the coupling of perception and action) are learned, but they also claim that this learning process did not evolve beyond a domain-general, all-purpose associative learning mechanism. Although we are strongly in favor of the first, more general associative argument (which is also clearly supported by Cook et al.'s review of the scientific evidence), we find their second claim less convincing, theoretically problematic (in terms of evolutionary theory), and somewhat counterproductive. Instead, we shall argue that a more realistic and productive associative account of MNs (and of similar advanced cognitive traits) should be based on identifying the genetically variable mechanistic details on which the associative process is based, and on which selection can operate in the course of evolution.
Cook et al.'s second claim that precludes genetic evolution is not as parsimonious as it may first appear. It implies that associative learning mechanisms acting in the construction of MNs for thousands of generations were not affected by selection in any way that was related to their role in producing MNs. This argument is difficult to test and is also evolutionarily unlikely. Unless it is clear that a trait is fairly recent (such as reading, typing, driving a car, etc.), a possible effect of natural selection on this trait cannot be precluded and is even to be expected. For example, although it is possible that our ability to use our hands for handling tools is merely a by-product of their adaptive design for climbing trees, we cannot preclude the possibility that our hands were also modified by natural selection as a result of their tool-making activity during the past ~3 million years. Moreover, the functionality and the conserved nature of associative learning mechanisms (highlighted by Cook et al.), suggest that they are under strong stabilizing selection. This stabilizing selection, however, is unlikely to be identical across multiple species and populations. It is both plausible and supported empirically that despite general similarity, associative learning mechanisms may differ in their parameters or mechanistic details as a result of different selection pressures. Many forms of learning, from imprinting to taste aversion, are increasingly recognized as different forms of associative learning that have adapted to different tasks (e.g., Bateson Reference Bateson1990; Shettleworth Reference Shettleworth2010). There is also evidence for fine-tuning of associative learning mechanisms by different selection pressures imposed by alternative behavioral strategies (Mery et al. Reference Mery, Belay, So, Sokolowski and Kawecki2007). Why, then, should associative learning mechanisms acting in the construction of MNs be different?
Paradoxically, the extreme version of the associative learning account weakens our ability to use associative learning to explain cognitive phenomena. If indeed all it takes to construct MNs (or similar associative networks) is to have the same basic domain-general associative learning ability, then all animals and all individuals should construct such networks equally well. Ignoring for the moment whether this is the case, adopting this view implies that heritable variation or species-specific differences in cognitive abilities cannot be explained by the associative account. Thus, the extreme version of the associative learning account may still explain MNs but cannot explain heritable variation in intelligence or social skills (Baron-Cohen et al. Reference Baron-Cohen, Richler, Bisarya, Gurunathan and Wheelwright2003; Gray & Thompson Reference Gray and Thompson2004), cannot address genetically based cognitive disorders (e.g., Crespi et al. Reference Crespi, Stead and Elliot2010), nor can it be used as a framework for cognitive evolution.
Alternatively, allowing associative learning mechanisms to vary genetically may offer a much more powerful account. We have recently proposed such a framework (Lotem & Halpern Reference Lotem and Halpern2012) based on earlier work (Goldstein et al. Reference Goldstein, Waterfall, Lotem, Halpern, Schwade, Onnis and Edelman2010; Lotem & Halpern Reference Lotem and Halpern2008) and their recent implementation (Kolodny et al. in press; 2014). In our framework, MNs may be viewed as no more than a specific instance of a much wider set of associative networks that represent contingencies or contiguities in time and space, which is consistent with Cook et al.'s general view. However, contrary to Cook et al., we predict that different species or individuals may construct different associative networks as a result of genetic differences in (a) their data-acquisition mechanisms (the attentional and motivational mechanisms directing them to process the relevant data) and in (b) the memory parameters (weight increase and decrease) of their associative learning rules. Most importantly, our model emphasizes the coevolved coordination between these two genetically variable components, coordination that determines learning dynamics and therefore the content and the structure of the network. Our data-acquisition mechanism is in fact similar to Heyes's idea of “input mechanisms” by which associative learning may be tuned to acquire data about the actions of others in order to facilitate social learning (Heyes Reference Heyes2012c). But we also go a step further by proposing that acquiring the relevant data is not sufficient. Memory parameters possessed by the learner must also fit the distribution of acquired data, as they presumably evolved to test the statistical significance of patterns and associations, given the natural distribution of the data. We used these principles to explain a range of phenomena in language acquisition and cognitive development (Lotem & Halpern Reference Lotem and Halpern2008; Goldstein et al. Reference Goldstein, Waterfall, Lotem, Halpern, Schwade, Onnis and Edelman2010; Kolodny et al. in press) and to propose an associative framework for cognitive evolution (Lotem & Halpern Reference Lotem and Halpern2012; Kolodny et al. Reference Kolodny, Edelman and Lotem2014). Although our model may eventually turn out to be inaccurate or even incorrect, it certainly suggests that allowing associative principles to evolve is a much more useful exercise than dismissing their genetic evolution. We thus call the readers to adopt Cook et al.'s general endorsement of associative accounts (see also Heyes Reference Heyes2012b; Reference Heyes2012c), but not to the point of neglecting the genetic evolution of their underlying mechanisms.
In promoting the associative learning account of mirror neurons (MNs), Cook et al. have chosen to adopt an extreme version. Not only have they argued that the matching properties of MNs (the coupling of perception and action) are learned, but they also claim that this learning process did not evolve beyond a domain-general, all-purpose associative learning mechanism. Although we are strongly in favor of the first, more general associative argument (which is also clearly supported by Cook et al.'s review of the scientific evidence), we find their second claim less convincing, theoretically problematic (in terms of evolutionary theory), and somewhat counterproductive. Instead, we shall argue that a more realistic and productive associative account of MNs (and of similar advanced cognitive traits) should be based on identifying the genetically variable mechanistic details on which the associative process is based, and on which selection can operate in the course of evolution.
Cook et al.'s second claim that precludes genetic evolution is not as parsimonious as it may first appear. It implies that associative learning mechanisms acting in the construction of MNs for thousands of generations were not affected by selection in any way that was related to their role in producing MNs. This argument is difficult to test and is also evolutionarily unlikely. Unless it is clear that a trait is fairly recent (such as reading, typing, driving a car, etc.), a possible effect of natural selection on this trait cannot be precluded and is even to be expected. For example, although it is possible that our ability to use our hands for handling tools is merely a by-product of their adaptive design for climbing trees, we cannot preclude the possibility that our hands were also modified by natural selection as a result of their tool-making activity during the past ~3 million years. Moreover, the functionality and the conserved nature of associative learning mechanisms (highlighted by Cook et al.), suggest that they are under strong stabilizing selection. This stabilizing selection, however, is unlikely to be identical across multiple species and populations. It is both plausible and supported empirically that despite general similarity, associative learning mechanisms may differ in their parameters or mechanistic details as a result of different selection pressures. Many forms of learning, from imprinting to taste aversion, are increasingly recognized as different forms of associative learning that have adapted to different tasks (e.g., Bateson Reference Bateson1990; Shettleworth Reference Shettleworth2010). There is also evidence for fine-tuning of associative learning mechanisms by different selection pressures imposed by alternative behavioral strategies (Mery et al. Reference Mery, Belay, So, Sokolowski and Kawecki2007). Why, then, should associative learning mechanisms acting in the construction of MNs be different?
Paradoxically, the extreme version of the associative learning account weakens our ability to use associative learning to explain cognitive phenomena. If indeed all it takes to construct MNs (or similar associative networks) is to have the same basic domain-general associative learning ability, then all animals and all individuals should construct such networks equally well. Ignoring for the moment whether this is the case, adopting this view implies that heritable variation or species-specific differences in cognitive abilities cannot be explained by the associative account. Thus, the extreme version of the associative learning account may still explain MNs but cannot explain heritable variation in intelligence or social skills (Baron-Cohen et al. Reference Baron-Cohen, Richler, Bisarya, Gurunathan and Wheelwright2003; Gray & Thompson Reference Gray and Thompson2004), cannot address genetically based cognitive disorders (e.g., Crespi et al. Reference Crespi, Stead and Elliot2010), nor can it be used as a framework for cognitive evolution.
Alternatively, allowing associative learning mechanisms to vary genetically may offer a much more powerful account. We have recently proposed such a framework (Lotem & Halpern Reference Lotem and Halpern2012) based on earlier work (Goldstein et al. Reference Goldstein, Waterfall, Lotem, Halpern, Schwade, Onnis and Edelman2010; Lotem & Halpern Reference Lotem and Halpern2008) and their recent implementation (Kolodny et al. in press; 2014). In our framework, MNs may be viewed as no more than a specific instance of a much wider set of associative networks that represent contingencies or contiguities in time and space, which is consistent with Cook et al.'s general view. However, contrary to Cook et al., we predict that different species or individuals may construct different associative networks as a result of genetic differences in (a) their data-acquisition mechanisms (the attentional and motivational mechanisms directing them to process the relevant data) and in (b) the memory parameters (weight increase and decrease) of their associative learning rules. Most importantly, our model emphasizes the coevolved coordination between these two genetically variable components, coordination that determines learning dynamics and therefore the content and the structure of the network. Our data-acquisition mechanism is in fact similar to Heyes's idea of “input mechanisms” by which associative learning may be tuned to acquire data about the actions of others in order to facilitate social learning (Heyes Reference Heyes2012c). But we also go a step further by proposing that acquiring the relevant data is not sufficient. Memory parameters possessed by the learner must also fit the distribution of acquired data, as they presumably evolved to test the statistical significance of patterns and associations, given the natural distribution of the data. We used these principles to explain a range of phenomena in language acquisition and cognitive development (Lotem & Halpern Reference Lotem and Halpern2008; Goldstein et al. Reference Goldstein, Waterfall, Lotem, Halpern, Schwade, Onnis and Edelman2010; Kolodny et al. in press) and to propose an associative framework for cognitive evolution (Lotem & Halpern Reference Lotem and Halpern2012; Kolodny et al. Reference Kolodny, Edelman and Lotem2014). Although our model may eventually turn out to be inaccurate or even incorrect, it certainly suggests that allowing associative principles to evolve is a much more useful exercise than dismissing their genetic evolution. We thus call the readers to adopt Cook et al.'s general endorsement of associative accounts (see also Heyes Reference Heyes2012b; Reference Heyes2012c), but not to the point of neglecting the genetic evolution of their underlying mechanisms.